Adding loops to a trained model has an inference cost

The cost comparison

The authors compare looped models against the same model trained without loops and against alternatives using the same inference computation. Reusing layers leaves parameter count unchanged but adds work for each generated token.

Adding loops to a trained model has an inference cost ↗

Where it helped in these tests

Across tested Qwen3 and Qwen3.5 models, only the 0.8-billion-parameter looped model beat both comparisons. The 4B and 9B cases did not beat the equally trained unlooped model. This finding concerns loops added after pretraining, not every form of recurrent model training.

Adding loops to a trained model has an inference cost ↗

What you can do

1. Compare equal-compute baselines before treating unchanged model size as unchanged inference cost. Open reference ↗

Where the evidence stops

Original abstract read November 4. Results are author-reported and were not independently reproduced; deployment latency and hardware costs were not measured.

Sources read. Primary sources read November 4, 2026 UTC. No installation, model inference or complete workflow executed.

Sources & further reading